How to pass CCAR-F: a four-week study plan

A week-by-week plan for the Claude Certified Architect - Foundations exam that follows the domain weights, builds one thing per domain, and uses the question bank as a diagnostic rather than a memory test.

Four weeks is enough for CCAR-F if you already build with Claude and can give it five or six hours a week. Double the plan if you are newer to the Agent SDK, Claude Code and MCP; the exam tests details that only stick once you have typed them.

The plan follows three rules. Study in proportion to the domain weights. Build one small thing per domain, because this exam tests exact names and values. And use the question bank to find gaps, not to memorise answers; the real exam draws four scenarios you cannot predict.

Before you start: two hours

Read the CCAR-F exam guide page once, including the six scenarios and the list of ruled-out topics. The ruled-out list saves more time than anything else on this plan: no fine-tuning, no embeddings, no streaming, no pricing maths, no cloud-provider configuration.

Then take the seven-question sample on that page, and note which domains felt like guessing.

Week 1: Domain 1, Agentic Architecture (27%)

The biggest domain and the one the others depend on. Read the six lessons in Domain 1 of the study guide, one a day.

Build: the smallest possible agent with the Agent SDK. One tool, a loop that branches on stop_reason, and a PostToolUse hook that changes a tool result before the model sees it. Then add a second agent type and spawn it with the Task tool; watch what happens when "Task" is missing from allowedTools.

Finish the week with the Domain 1 quiz in the question bank. Under 80%? Reread the lessons whose task statements you missed; the quiz shows them.

Week 2: Domain 2, Tools and MCP (18%), and Domain 3, Claude Code (20%)

Two domains, both detail-heavy. Read Domain 2 in the first half of the week and Domain 3 in the second.

Build for Domain 2: write two tools with deliberately similar jobs, give them one-line descriptions, and watch Claude misroute. Rewrite the descriptions with inputs, examples and boundaries and watch it stop. Return one error with isError: true and structured metadata.

Build for Domain 3: on a real project, set up CLAUDE.md at project level, add one .claude/rules/ file with a paths glob, one command in .claude/commands/, one skill with context: fork and allowed-tools, and run claude -p "..." --output-format json from a script. Run /memory and /compact once each so you have seen them.

Both domain quizzes at the end of the week.

Week 3: Domain 4, Prompting and Structured Output (20%), and Domain 5, Context and Reliability (15%)

Read Domain 4 and Domain 5.

Build for Domain 4: an extraction tool with a JSON schema, forced with tool_choice, with one optional field and one enum that has an "other" value. Feed it a document missing the optional field and confirm you get null, not an invention. Send one small job through the Message Batches API and match the results by custom_id.

Build for Domain 5: nothing to build; instead, write the case-facts block you would keep for a support conversation, and the structured error a subagent should return on a timeout. Writing them out fixes the shapes in memory.

Both domain quizzes, then the cheat sheet read end to end.

Week 4: simulate, review, repeat

Sit the 60-question timed simulator in the question bank on day one of the week, at full length, no notes. Whatever the score, read every explanation, including for the questions you got right. The explanations for the wrong options are where the exam's distractors live.

Spend the middle of the week on the two weakest domains from the simulator report. Reread those lessons; redo those domain quizzes.

Sit a second simulator near the end of the week. Treat anything under 75% as "not yet", because the real exam draws four of six scenarios and your four may be the harder ones. At 80% or above on two simulators, book.

The night before

Reread the cheat sheet. Check the exam-day rules on the exam page: government ID matching your registration name, a clear desk, no second monitor, the Pearson system test if you are testing online. Sleep.

If you have less than four weeks

Cut the builds before you cut the reading, but keep the Domain 1 build; the loop, hooks and Task tool are where the most questions are. Cut Week 4 to one simulator. Do not cut the ruled-out list.

Written by Anton Pavelko, who holds the Claude Certified Architect, Associate and Developer credentials. An independent resource, not affiliated with Anthropic. Spotted an error? Email a correction.